using!lasso!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε...

27
Sandia National Laboratories is a multi-program laboratory managed and operated by Sandia Corporation, a wholly owned subsidiary of Lockheed Martin Corporation, for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-AC04-94AL85000. SAND2014-2429C Using LASSO to infer a highorder eddy viscosity model for kε RANS simula=on of transonic flows S. Lefantzi, J. Ray, S. Arunajatesan and L. Dechant ([email protected]) SAND2016-4661C

Upload: others

Post on 20-Jan-2020

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

Sandia National Laboratories is a multi-program laboratory managed and operated by Sandia Corporation, a wholly owned subsidiary of Lockheed Martin Corporation, for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-AC04-94AL85000. SAND2014-2429C

 Using  LASSO  to  infer  a  high-­‐order  eddy  viscosity  model  for  k-­‐ε  

RANS  simula=on  of  transonic  flows    S. Lefantzi, J. Ray, S. Arunajatesan and L. Dechant

([email protected])  

SAND2016-4661C

Page 2: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

The  problem  

§  Aim:  Develop  a  predic=ve  k-­‐ε    RANS  model  for  transonic  jet-­‐in-­‐crossflow  (JinC)  simula=ons  

§  Drawback:  RANS  simula=ons  are  simply  not  predic=ve  §  They  have  “model-­‐form”  error  i.e.,  missing  physics    §  The  numerical  constants/parameters  in  the  k-­‐ε  model  are  usually  

derived  from  canonical  flows    

§  Hypothesis    §  One  can  calibrate  RANS  to  jet-­‐in-­‐crossflow  experiments;  thereaRer  the  

residual  error  is  mostly  model-­‐form  error  §  Due  to  model-­‐form  error  and  limited  experimental  measurements,  the  

parameter  es=mates  will  be  approximate  §  We  will  es=mate  parameters  as  probability  density  func=ons  (PDF)  

§  We  then  address  the  model-­‐form  error  with  an  enriched  eddy  viscosity  model  for  the  missing  physics  

2  

Page 3: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

The  equa=ons  §  The  model  

§  Devising  a  method  to  calibrate  k-­‐ε  parameters  from  expt.  data  

 

 

 

§  Sources  of  errors  §  Parameters  {C2,    C1}    are  obtained  from  canonical  flows    §  Cµ  is  deemed  constant  throughout  the  flowfield  §  Linear  stress-­‐strain  rate  rela=onship    τij  =  -­‐2/3k  δij  +  µT  Sij  

§  Called  a  linear  eddy  viscosity  model  (LEVM)  3  

∂ρk∂t

+∂∂xi

ρuik − µ +µTσ k

#

$%

&

'(∂k∂xi

)

*+

,

-.= Pk − ρε + Sk

∂ρε∂t

+∂∂xi

ρuiε − µ +µTσε

#

$%

&

'(∂ε∂xi

)

*+

,

-.=

εkC1 f1Pk −C2 f2ρε( )+ Sε

µT =Cµ fµρk2

ε

Page 4: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

Target  problem  -­‐  jet-­‐in-­‐crossflow  

§  A  canonical  problem  for  spin-­‐rocket  maneuvering,  fuel-­‐air  mixing  etc.  

§  We  have  experimental  data  (PIV  measurements)  on  the  cross-­‐  and  mid-­‐plane  

§  Will  calibrate  to  vor=city  on  the  crossplane  and  test  against  mid-­‐plane  

4  −0.06−0.04−0.0200.020.040.06

0

0.05

0.1

Z (m)

Y (m

)

−4000

−3000

−2000

−1000

0

1000

2000

3000

4000

5000

Page 5: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

RANS  (k-­‐ω)  simula=ons  -­‐  crossplane  results  

§  Crossplane  results  for  stream  §  Computa=onal  results  (SST)  are  too  round;  Kw98  doesn’t  have  

the  mushroom  shape;  non-­‐symmetric!  §  Less  intense  regions;  boundary  layer  too  weak  

5  

Page 6: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

Reducing  errors    §  Model-­‐form  errors  

§  The  linear  turbulent  stress  –  strain  rate  rela=onship  (LEVM)  can  be  enriched  with  quadra=c  and  cubic  terms  (QEVM  /  CEVM)  §  Includes  terms  with  vor=city  and  cross  terms  with  vor=city  and  strain  rate  

§  However  the  high-­‐order  models  have  parameters  in  them  §  What  are  the  appropriate  values  for  those  parameters?  

§  Parametric  uncertainty  §  (C2,  C1)  can  be  es=mated  (somewhat)  from  experimental  data  

§  But  because  of  model-­‐form  errors  and  limited  experimental  data,  these  cannot  be  es=mated  with  much  certainty  

§  We’ll  use  Bayesian  inversion  and  es=mate  them  as  PDFs  §  Quan=fies  uncertainty  in  the  es=mate  of  the  parameters  

§  Calibra=on  process  §  Iden=fy  which  of  the  CEVM  parameters  can  actually  be  es=mated  

from  experimental  data  §  Then  calibrate  those  along  with  (C2,  C1);  call  the  full  set  C  =  (:,  C2,  C1)  

6  

Page 7: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

Calibra=on  details  §  Aims  of  the  calibra=on  

§  Calibrate  to  a  M  =  0.8,  J  =  10.2  interac=on  §  Learn  the  form  of  the  high-­‐order  eddy  viscosity  model  by  fiong  to  

turbulent  stresses  measurements  on  the  mid-­‐plane  §  Calibrate  to  crossplane  data;  check  by    matching  the  midplane  velocity  

profiles  

§  Technical  challenges  §  Computa=onal  cost  of  3D  JinC  RANS  simula=on    

§  Replace  3D  RANS  with  a  surrogate  model  i.e.,  model  crossplane  streamwise  vor=city  ω(RANS)

x(y)  =  f(y;  C),  f(:;  C)  is  a  curve-­‐fit  §  Surrogate  model  =  emulators  

§  Arbitrary  combina=ons  of  C  may  be  nonphysical  §  How  to  build  emulators  when  C  are  nonsensical?  

§  What  func=onal  form  to  use  for  f(:;  C)?  

Page 8: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

High-­‐order  eddy-­‐viscosity  model  

§  CraR  95  describes  a  cubic  eddy  viscosity  (CEVM)  model  §  τij  =  -­‐2/3k  δij  +  CµF(Sij,  ε)  +  c1f1(Sij,  Ωij,  ε)  +  c2f2(Sij,  Ωij,ε)  …..  c7f7(Sij,  Ωij, ε)  §  F(Sij)  is  linear  in  Sij,  f1(:,  :,  :)  -­‐  f3(:,  :,  :)  are  quadra=c  in  Sij  &  Ωij  

§  f4(:,  :,  :)  –  f7(:,  :,  :)  are  cubic  in  Sij  &  Ωij  

§   Our  experimental  data,  on  the  midplane,  consists  of:  §  Sij  &  Ωij  obtained  from  the  measured  velocity  field  §  τij  and  k,  also  measured  §  ε  (dissipa=on  rate  of  turbulent  KE)  cannot  be  measured  

§  It  is  approximated  by  assuming  equilibrium  of  produc=on  and  dissipa=on  of  turbulent  KE.  

§  CraR’s  model  prescribes  {c1  …  c7}  §  Parameter  value  obtained  from  a  simple,  incompressible  turning  flow  §  May  not  be  valid  for  transonic  JinC  interac=on  

8  

Page 9: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

Es=ma=on  of  CEVM  parameters  

§  The  180  measurements  that  we  have  may  not  have  info  that  informs  c1  …  c7  

§  Cast  the  es=ma=on  problem  as    

§  The  first  half  es=mates  x  =  {ci}  that  provide  CEVM  predic=ons  near  Y  §  The  second  half  –  the  λ  penalty  –  tries  to  set  as  many  ci  to  zero  §  Called  Shrinkage  Regression  

§  The  penalty  λ  is  the  lynchpin  §  If  it  is  too  small,  we  get  over-­‐fiong  (too  many  ci  survive)  §  The  best  way  to  get  λ  is  via  k-­‐fold  cross-­‐valida=on  

§  The  method  for  solving  the  op=miza=on  problem  is  LASSO   9  

minx

Y − Ac2

2+λ c

1

Page 10: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

k-­‐fold  cross-­‐valida=on  §  Divide  the  180  measurements  into  8  “folds”  (equal  subsets)  §  Pick  a  value  of  λ’

§  Pick  fold  #  1  as  the  tes=ng  set,  folds  2-­‐8  as  the  learning  set  §  Solve  the  op=miza=on  problem  (solve  for  c)  using  Y  constructed  from  

the  learning  set  §  Predict  the  data  in  the  tes=ng  set  §  Repeat  with  folds  #2,  #3  …  as  the  tes=ng  sets  §  Obtain  the  mean  error  and  error  bars  for  λ’  

§  Ul=mately  you  get  error  as  a  func=on  of  λ§  Pick  the λ with  min  error  

§  For  higher  values  of  λ,  expect  to  see  lots  of  ci  becoming  zero  §  And  predic=ve  errors  becoming  large  

§  Nomenclature:  The  norm  of  difference  (Y(obs)  –  Ac)  is  called  the  ‘deviance’   10  

Page 11: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

LASSO  results  

§  CraR  explains  around  28%  of  deviance  §  As  log(λ)  increases  and  #  of  terms  retained  decreases,  CEVM  worsens  §  One  gets  λmin  and  λ1se  

Looking for a good λ

log(λ) log(λ)

Rel

ativ

e D

evia

nce

No.

of c

oeffi

cien

ts

log(λ) M

ean-

Squa

red

Erro

r

Page 12: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

Tabulate  coefficients  and  MSE  

Method c1 c2 c3 c4 c5 c6 c7 MSE Craft -0.1 0.1 0.26 -10 0 -5 5 0.662 λmin -0.065 -0.103 1.68 -4.02 5.7 5.4 -3.64 0.386 λ1se 0.0 0.0 0.455 0.0 0 0 0 0.483 LM -0.0789 -0.149 2.02 -5.88 0 6.68 -11.87 0.382

§  ln(λmin)      =  -­‐5.11,  ln(λ1se)  =  -­‐1.75  §  CraR’s  default  parameters  are  changed  when  we  regress  it  to  data  

§  Results  called  ‘LM’  

§  When  we  LASSO  the  model  using  λ1se,  we’re  leR  with  just  1  quadra=c  term  §  But  the  model  loses  much  accuracy  

§  Let’s  choose  λ1se.    §  Provides  a  simple  model,  and  keeps  the  Ω2  term  

Page 13: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

Calibra=on  of  {c3,  C2,  C1}  §  We  will  calibrate  C  =  (c3,  C2,  C1)  

§  Our  model  really  has  a  quadra=c  eddy  viscosity  model  (QEVM)  

§  Approach:  §  Data:  Use  vor=city  measurements  on  crossplane  to  es=mate  C  

§  Useful  measurements  available  at  225  loca=ons  (“probes”)  

§  Es=ma=on  procedure:  Bayesian  calibra=on  using  MCMC  §  Model:  Use  surrogate  models  (emulators)  of  the  RANS  simulator  

§  Set  of  1275  runs  in  the  parameter  space  C to  make  the  training  data  §  Iden=fy  a  physically  realis=c  space  R, use  SVMs  to  model  R §  Make  emulators  ω(C)  =  f(c3,  C2,  C1)  with  polynomials;  valid  in  R §  Use  MCMC  to  create  the  posterior  PDF  of  C  

§  Checking  results  §  Draw  100  samples  from  the  posterior  PDF  §  Develop  an  ensemble  of  predic=ons  of  vor=city  and  velocity;  compare  against  measurements  

13  

Page 14: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

The  Bayesian  calibra=on  problem  • Model experimental values at probe j as ω(j)

ex = ω(j)(C) + ε(j), ε(j) ~ N(0, σ2)

• Given prior beliefs π on C, the posterior density (‘the PDF’) is • P(C|ωex) is a complicated distribution that has to be described/

visualized by drawing samples from it • This is done by MCMC

–  MCMC describes a random walk in the parameter space to identify good parameter combination

–  Each step of the walk requires a model run to check out the new parameter combination

Λ ωex( j ) |C( )∝ exp −

ωex( j ) −ω ( j ) (C)( )

2

2σ 2

$

%

&&

'

(

))

j∈P∏

P(C,σ |ωex( j ) )∝Λ(ωex

( j ) |C,σ ) π (c3,C2,C1) πσ (σ )

Page 15: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

Making  emulators  -­‐  1  

§  Training  data  §  Sample  the  parameter  space  C  =  {c3,  C2,  C1};  bounds  are  known  §  Run  RANS  models  at  1275  samples;  save  vor=city  on  cross-­‐plane  §  Select  the  top  25%  of  the  training  runs  

§  Call  this  subspace  of    R §  Keeps  us  out  of  non-­‐physical  parts  of  the  parameter  space  C

§  Making  emulators  in  R §  Model  vor=city  at  probe  j  ω(j)  as  a  polynomial  in  C  

§  Simplify  using  AIC;  cross  validated  using  repeated  random  sub-­‐sampling  (100  rounds)  §  RMSE  in  Learning  &  Tes=ng  sets  should  be  equal  

§  Accept  all  surrogate  models  that  have  <  10%  error  15  

ω ( j ) ≅ a0 + a1c3 + a2C2 + a3C1 + a4c3C2 + a5c3C1 + a6C2C1 +.....

Page 16: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

Making  emulators  -­‐  2  

§  Emulators  with  10%  accuracy  could  only  be  made  for  55  /  224  probes    §  90  with  large  vor=city  

(circles)  §  55  with  emulators  (+)  

§  Also,  the  emulators  are  only  applicable  in  the  R    sec=on  of  the  parameter  space  C

Page 17: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

Making  the  informa=ve  prior    §  Our  emulators  are  valid  only  inside  R  

in  the  parameter  space C §  During  the  op=miza=on  (MCMC)  we  

have  to  reject  parameter  combina=ons  outside  R  (this  is  our  prior  belief  πprior(C))  §  We  define  ζ(C)  =  1,  for  C  in  R  and  ζ(C)  

=  -­‐1  for  C  outside  R §  Then  the  level  set  ζ(C)  =  0  is  the  

boundary  of  R §  The  training  set  of  RANS  runs  is  used  

to  populate  ζ(C)  §  We  have  to  “learn”  the  discrimina=ng  

func=on  ζ(C)  =  0    §  We  do  that  using  support  vector  

machine  (SVM)  classifiers  17  

Runs in the top 25th percentile

0.06 0.07 0.08 0.09 0.10 0.11 0.12

1.20

1.25

1.30

1.35

1.40

1.45

1.50

1.55

1.71.8

1.92.0

2.1

C2

C1

Page 18: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

PDFs  from  the  calibra=on      

§  About  60,000  MCMC  steps  to  convergence  

§  Calibrated  values  of  C  quite  different  from  the  ones  from  literature  §  Ver=cal  lines  are  the  

“canonical”  values  of  the  parameters  

§  Next  step  §  Draw  100  samples  from  

the  posterior  distribu=on  and  perform  RANS  simula=ons  

§  Compare  with  experimental  measurements  

Page 19: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

QEVM  point  vortex  metrics  §  Compare  measured  and  

simulated  vor=city  fields  using  the  circula=on,  the  centroid  and  radius  of  gyra=on  of  the  vor=city  distribu=on  §  Called  the  “point  vortex  

metrics”  

§  Comparable  results  using  exis=ng  LEVM  models  have  20%-­‐70%  errors  

Page 20: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

QEVM  PPT  predic=ons  on  midplane  

§  Use  the  100  RANS  simula=ons  to  obtain  velocity  field  on  the  mid-­‐plane  

§  Compare  experimental  and  simulated  predic=ons  

Page 21: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

Conclusions  §  We  are  beginning  to  “fix-­‐up”  engineering  models  with  observa=onal  data  

§  Includes  both  es=ma=ng  model  parameters  and  enriching  closure  models  (inferring  missing  physics  in  models)  

§  Methods  are  Bayesian;  fully  probabilis=c  inference  (of  parameters,  at  least)  §  Accommodates  uncertainty  in  es=mates  due  to  limited  data  and  shortcomings  of  

the  RANS  model  (model-­‐form  error)  

§  We  can  tackle  rather  complicated  problems  using  Bayesian  inference  §  Computa=onal  costs  are  immense,  but  only  for  genera=ng  training  data    §  Bri|le  –  we  depend  on  emulators,  which  can’t  always  be  made  §  Can  tackle  peculiari=es  of  non-­‐physical  parameter  spaces  using  informa=ve  

priors  (classifiers)  §  Tools  and  theories:  A  mixture  of  sta=s=cs  and  machine  learning  

§  Bayesian  inference,  emulators,  shrinkage  are  conven=onally  sta=s=cal  §  Classifiers  etc.  are  purely  ML  §  As  we  scale  up  and  confront  large  data  (simulated  flowfields  etc.)  to  infer  

model-­‐form  error,  expect  MapReduce  implementa=ons  of  these  tools  21  

Page 22: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

BONEYARD  

22  

Page 23: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

RANS  (k-­‐ω)  simula=ons  –  midplane  results  

§  Experimental  results  in  black  §  All  models  are  pre|y  inaccurate  (blue  and  red  lines  are  the  non-­‐

symmetric  results)  

U-­‐defect   V  -­‐  velocity  

Page 24: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

What  is  MCMC?  §  A  way  of  sampling  from  an  arbitrary  distribu=on  

§  The  samples,  if  histogrammed,  recover    the  distribu=on  

§  Efficient  and  adap=ve  §  Given  a  star=ng  point  (1  sample),  the  MCMC  chain  will  sequen=ally  

find  the  peaks  and  valleys  in  the  distribu=on  and  sample  propor=onally  

§  Ergodic  §  Guaranteed  that  samples  will  be  taken  from  the  en=re  range  of  the  

distribu=on  

§  Drawback  §  Genera=ng  each  sample  requires  one  to  evaluate  the  expression  for  

the  density  π§  Not  a  good  idea  if  π  involves  evalua=ng  a  computa=onally  expensive  

model  

Page 25: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

An  example,  using  MCMC  §  Given:  (Yobs,  X),  a  bunch  of  n  observa=ons  §  Believed:  y  =  ax  +  b  §  Model:  yiobs  =  axi  +  bi  +  εi,  ε  ~  N(0,  σ)  §  We  also  know  a  range  where  a,  b  and  σ  might  lie  

§  i.e.  we  will  use  uniform  distribu=ons  as  prior  beliefs  for  a,  b,  σ

§  For  a  given  value  of  (a,  b,  σ),  compute  “error”  εi  =  yiobs  –  (axi  +  bi)  §  Probability  of  the  set  (a,  b,  σ)  =    Π  exp(  -­‐  εi2/σ2  )  

§  Solu=on:  π  (  a,  b,  σ  |  Yobs,  X  )  =  Π  exp(  -­‐  εi2/σ2  )  *  (bunch  of  uniform  priors)  §  Solu=on  method:  

§  Sample  from  π  (  a,  b,  σ  |  Yobs,  X  )  using  MCMC;  save  them  §  Generate  a  “3D  histogram”  from  the  samples  to  determine  which  region  in  the  

(a,  b,  σ)  space  gives  best  fit    §  Histogram  values  of  a,  b  and  σ,  to  get  individual  PDFs  for  them  §  Es=ma=on  of  model  parameters,  with  confidence  intervals!  

Page 26: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

MCMC,  pictorially  §  Choose  a  star=ng  point,  Pn  =  

(acurr,  bcurr)  §  Propose  a  new  a,  aprop  ~  

N(acurr,  σa)  §  Evaluate  π  (  aprop,  bcurr  |  ...)  /  π  (  acurr,  bcurr  |  …  )  =  m    

§  Accept  aprop  (i.e.  acurr  <-­‐  aprop)  with  probability  min(1,  m)  

§  Repeat  with  b  §  Loop  over  =ll  you  have  

enough  samples  a

a

b

b

a

Proposal distribution

“good” values of (a, b)

Page 27: Using!LASSO!to!infer!ahigh2order!eddy!viscosity!model!for!k2 ε …jairay/Presentations/sand2016... · 2016-05-16 · Sandia National Laboratories is a multi-program laboratory managed

What  is  a  SVM  classifier?  §  Given  a  binary  func=on  y  =  f(x)  as  a  

set  of  points  (yi,  xi),  yi  =  (0,  1)  §  Find  the  hyperplane  y  +  Ax  =  0  that  

separates  the  x-­‐space  into  y  =  0  and  y  =  1  parts  

§  Posed  as  an  op=miza=on  problem  that  maximizes  the  margin  

27  

§  In  case  of  a  curved  discriminator,  need  a  transforma=on  first  §  Achieved  using  kernels  §  We  use  a  cubic  kernel